Comparison of Several Simplistic High-Level Approaches for Estimating the Global Energy and Electricity Use of ICT Networks and Data Centers

Abstract

International Journal of Green Technology 2019; 5(1): 50-63.
Currently the global energy and electricity use of ICT networks and data centers are estimated and predicted by several different top-down approaches. It has not been investigated which prediction approach best answers to the 5G, Artificial Intelligence and Internet of Things megatrends which are expected to emerge until 2030 and beyond. The analysis of the potential correlation between storage volume, communication volume and computations (instructions, operations, bits) is also lacking. The present research shows that several different activity metrics (AM)-e.g. data traffic, subscribers, capita, operations-have and can be been used. First the global baseline electricity evolution (TWh) for 2010, 2015 and 2020 for networks of fixed, mobile and data centers is set based on literature. Then the respective AM-e.g. data traffic-associated with each network are identified. Then the following are proposed: Compound Aggregated Growth Rate (CAGR) for each AM, CAGR for TWh/AM and the resulting TWh values for 2025 and 2030. The results show that AMs based on data traffic are best suited for predicting future TWh usage of networks. Data traffic is a more robust (scientific) AM to be used for prediction than subscribers as the latter is a more variable and less definable concept. Nevertheless, subscriber based AM are more uncertain than data traffic AM as the subscriber is neither a well-defined unit, nor related to the network equipment which handle the data. Despite large non-chaotic uncertainties, data traffic is a better AM than subscribers for expressing the energy evolution of ICT Networks and Data Centers. Top-down/high-level models based on data traffic are sensitive to the amount of traffic however also to the development of future electricity intensity. For the first time the primary energy use of computing, resulting from total global instructions and energy per instruction, is estimated. Combining all networks and data centers and using one AM for all does not reflect the evolution improvement of individual network types. Very simplistic high-level estimation models tend to both overestimate and underestimate the TWh. However, looking at networks and data centers as one big entity better reflects the future converging paradigm of telecom, ICT and computing. The next step is to make the prediction models more sophisticated by using equipment standards instead of top-down metrics. The links between individual equipment roadmaps (e.g. W/(bits per second)) and sector-level roadmaps need further study.

Full text

50 International Journal of Green Techno logy, 2019, 5, 50-63
E-ISSN: 2414-2077/19 © 2019 International Journal of Green Technology
Comparison of Several Simplistic High-Level Approaches for
Estimating the Global Energy and Electricity Use of ICT Networks
and Data Centers
Anders S.G. Andrae*
Huawei Technologies Sweden AB, Skalholtsgatan 9, 16494 Kista, Sweden
Abstract: Currently the global energy and electricity use of ICT networks and data centers are estimated and predicted
by several different top-down approaches. It has not been investigated which prediction approach best answers to the
5G, Artificial Intelligence and Internet of Things megatrends which are expected to emerge until 2030 and beyond. The
analysis of the potential correlation between storage volume, communication volume and computations (instructions,
operations, bits) is also lacking. The present research shows that several different activity metrics (AM) – e.g. data traffic,
subscribers, capita, operations – have and can be been used. First the global baseline electricity evolution (TWh) for
2010, 2015 and 2020 for networks of fixed, mobile and data centers is set based on literature. Then the respective AM –
e.g. data traffic - associated with each network are identified. Then the following are proposed: Compound Aggregated
Growth Rate (CAGR) for each AM, CAGR for TWh/AM and the resulting TWh values for 2025 and 2030. The results
show that AMs based on data traffic are best suited for predicting future TWh usage of networks. Data traffic is a more
robust (scientific) AM to be used for prediction than subscribers as the latter is a more variable and less definable
concept. Nevertheless, subscriber based AM are more uncertain than data traffic AM as the subscriber is neither a well-
defined unit, nor related to the netw ork equipment which handle the data. Despite large non-chaotic uncertainties, data
traffic is a better AM than subscribers for expressing the energy evolution of ICT Networks and Data Centers. Top-
down/high-level models based on data traffic are sensitive to the amount of traffic however also to the development of
future electricity intensity. For the first time the primary energy use of computing, resulting from total global instructions
and energy per instruction, is estimated.
Combining all networks and data centers and using one AM for all does not reflect the evolution improvement of
individual network types. Very simplistic high-level estimation models tend to both overestimate and underestimate the
TWh. However, looking at networks and data centers as one big entity better reflects the future converging paradigm of
telecom, ICT and computing.
The next step is to make the prediction models more sophisticated by using equipment standards instead of top-down
metrics. The links between individual equipment roadmaps (e.g. W/(bits per second)) and sector-level roadmaps need
further study.
Keywords: ICT, data centers, data traffic, model, communication networks, electricity, prediction, primary energy,
subscribers.
1. INTRODUCTION
Electricity is at ≈44% (electricity production 26614.8
TWh/4.4 ≈ 6050 Million tonnes oil equivalents (Mtoe) of
13800 Mtoe) a huge share of the current global primary
energy supply (PES) [1]. The global PES grew by 16
EJ between 2017 and 2018 and 10 EJ between 2016
and 2017 [1]. No predictions are done in [1] which
reports factual observations, however, predictions can
still be attempted based on historical trends. With the
current growth rates [1], extrapolated between 2018
and 2030, global PES would rise from 13800 Mtoe to
16600 Mtoe, and the electricity supply (not PES) from
26000 to 36000 TWh. However, needless to point out,
it is important to always critically question forecasts [2].
One important driver for increasing global electricity
demand is for millions in poor underdeveloped
households to quickly get access to reliable electric
Address correspondence to this article at the Huawei Technologies Sweden
AB, Skalholts gatan 9, 16494 Kista, Sweden; Tel: +46-739-200-533; Fax: +46-
812-060-800; E-mail: anders.andrae@huawe i.com
power which can replace health damaging direct coke
and wood burning [3]. Other important drivers will be
electrification of transport and sustainable
hydrogen/ammonia production. Electricity is more
energy efficient for transport than using primary fuel [4].
Therefore the total global PES could be reduced by
using more electric power in the use stage. However,
the manufacturing of batteries to be used in electric
vehicles requires around 350-650 MJ PES/kWh battery
capacity [5], which corresponds to around 200 kWh
electric power/kWh. Moreover, until there is charging
infrastructure across the globe or batteries that can last
for ≈1000 km the World will be dependent on internal
combustion engines [6]. ICT is essential in society, the
market is expanding fast and there is some concern
that ICT might use many times more electric power in
the next decade [7]. The driving trends for data traffic
and computations are more advanced video, video
streaming for gaming, augmented reality, 5G, artificial
intelligence (AI) training, autonomous vehicles with
streaming cameras, holography, digitalization,
advanced commercials, and the need for reliable

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 51
electricity. 5G technology will foster more demand from
online gaming, video calling and interactive sports
experiences. If the current energy intensity J/bit (≈3-5
pJ/bit [8,9]) – expressed as PES/instruction - does not
improve, it is (unrealistically?) estimated that in 2030
the PES use of computing devices could reach ≈60%
of the total amount of global PES, i.e. ≈420 EJ. This will
become completely unsustainable by 2040 [9,10]. In
this research, total global instructions and total global
bits are considered equal. Anyway, the scientific
method requires clear definitions. Moreover, correct
judgement of changes and trends require reliable
observations and homogeneous and representative
measurements. Global ICT energy and electricity use
predictions have neither but still the order of magnitude
of the historical TWh electricity use seem commonly
agreed and understood. Understanding the trends for
energy and electricity demand need more than a
couple of years of factual observed data. Maybe
something like ten years is required in order to
understand if ICT and computing has changed its
electricity use.
When in operation, ICT Networks and data centers
use around at least 500 TWh electricity estimated by
several different approaches [7,18]. Consumer ICT
devices, embedded chips and production of ICT
equipment could add up another 1500 TWh [7] but this
is not explored further in the present research. It has
been proposed that training a single AI model -
Transformer (213M parameters) with neural
architecture search - will require around 660 MWh
electricity [11]. Andrae explored how much electricity
might be required toward 2030 under certain
circumstances which are not at all unrealistic
(exploding data due to 5G, Machine Learning and AI +
transistor energy problems) [12]. Andrae explored
Special Purpose Computing instructions. However, the
high-level system-level energy saving is somewhat
ignored in such approaches. Anyway, the PES intensity
per logic operation (J/bit) and the total global number of
logic operations (bits) have not been explored despite
work by e.g. Åberg and Mämmela [8] discussing
energy limited microelectronics.!!
Questions explored in the present research:
• Which simplistic high-level approach gives the
most credible and systematic predictions for
2030 ICT networks and data centers electricity
use in the use stage?
• What is the probability of the electricity use of
ICT networks and data centers rising 5-10 times
as has previously been suggested [7] ?
• Which models can best mirror what is in the ICT
network power calculation?
Table 1: Bases for understanding ICT Sector electricity footprint on a top-down/high-level.
Sections 2 and 3
Basis for prediction, Activity Metric (AM)
Reference
“Fixed”
Data traffic
[7]
“Fixed”
Data traffic
[13]
“Fixed”
“Fixed” subscribers
[14]
“Mobile”
Data traffic
[7]
“Mobile”
Data traffic
[13]
“Mobile”
“Mobile” subscribers
[14]
“Data Centers”
Data traffic
[7]
“Data Centers”
Data traffic
[13]
“Data Centers”
“Servers”
[15]
“Fixed”+ “Mobile” + “Data Centers”
Data traffic
This research
Section 4
“Data Centers”
Historical Compound Aggregated Growth Rate (CAGR) of
electricity (TWh)
[16]
“Fixed”+”Mobile”
Historical CAGR of electricity (TWh)
[16]
“Fixed”+ “Mobile” + “Data Centers”
CAGR of Global Capita and TWh/Global Capita
This research
“Fixed”+ “Mobile” + “Data Centers”
Bits, operations, computations, instructions
This research

52 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
• Which model helps us understand and predict
ICT network power?
• Has the electricity consumption by data centers
leveled off in recent years due to efficiency
improvements?
The originality of this research is the systematic
comparison of prediction results made possible by
several different models (Table 1).
The falsifiable hypotheses for models describing
ICT electricity use described in literature are:
• Data traffic numbers from [13] are – if used
carefully - the most suitable prediction bases
(activity metrics, AMs) for networks and data
centers electricity use
• Final electricity use (TWh) for all kind of
networks is equally sensitive to changes in Data
traffic growth and electricity intensity
improvement
2. MATERIALS AND METHODOLOGY
In this section it is explained which methods are
used to obtain the results shown in section 3.
Equations 1-3 describe the idealized relations
between electricity usage for each year from 2021 to
2030, activity metric, and electricity intensity
improvement. Electricity usage for 2010, 2015 and
2020, and corresponding activity metric (AM), and
resulting electricity intensity improvement between
2010 and 2020 are assumed known facts.
E2021+n=AM2021+n
AM 2020+n
!
"
#$
%
&'E2020+n'EI2010 to 2020
(1)
AM 2021+n=AM2020+n!(CAGRAM ,2020 to 2030 )n
(2)
EI2010 to 2020 =
E2020
AM 2020
!
"
#$
%
&
E2010
AM 2010
!
"
#
#
#
#
#
$
%
&
&
&
&
&
1
(2020'2010 )
(3)
where
E,2010 = electricity usage in networks and data centers
in 2010, TWh
E,2020 = electricity usage in networks and data centers
in 2020, TWh
E,2021 = electricity usage in networks and data centers
in 2021, TWh
AM2020 = Activity metric in 2020, e.g. Exabyte (EB),
subscriber etc.
AM2021 = Activity metric in 2021, e.g. EB, subscriber
etc.
CAGRAM,2020 to 2030
= Compound Aggregated Growth
Rate (CAGR) for Activity metric in between 2020 and
2030, %
n = 0,1,2,3…,9.
EI2010 to 2020
= annual electricity intensity improvement
between 2010 and 2020, %
EI2020 to 2030
= annual electricity intensity improvement
between 2020 and 2030, %
EI2010 to 2020 = EI2020 to 2030.
CAGRAM,2020 to 2030 might differ considerably for data
traffic in between literature sources and networks.
Table 2 shows baseline values for electricity and data
traffic footprint of Networks in 2010, 2015 and 2020.
624 TWh for “Fixed”+ “Mobile” + “Data Centers” for
2010 from [19] is obtained from backward extrapolation
of growth rates between 2013 and 2015.
207 TWh for data center global electricity seems
very low considering that it was recently estimated that
China’s data centers alone used 160 TWh [20].
Regarding “fixed” networks use stage,
Kyriakopoulos et al. demonstrated that for Elastic
Optical Networks Based on Signal Overlap there is an
almost linear increase in power consumption, when the
average data traffic demand increases [21].
Such estimations strengthen the arguments that
global data centers and “fixed” networks – despite
removing old “fixed” networks” - could use more
electricity in the next decade.
3. RESULTS
Here follows a concise and precise description of
the results. Table 3 shows the evolution of data traffic,
subscribers and TWh for “fixed”, “mobile” and “data
centers”.

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 53
Table 2: Baseline Values for Electricity and Data Traffic Footprint of Networks in 2010, 2015 and 2020
Network type
E (TWh) and Activity Metrics (AM)
2010
2015
2020
“Fixed”
Data traffic, ExaByte (EB) [7]
325
839
2444
“Fixed”
Data traffic, EB [13]
325
839
2568
“Fixed”
“Subscribers”, billions [14]
1.75
1.85
2
“Fixed”
Electricity use, TWh [17]
162
179
171
“Fixed”
Electricity use, TWh [18]
80
85
90
“Mobile”
Data traffic, EB [7]
22.6
75
791
“Mobile”
Data traffic, EB [13]
22.6
75
492
“Mobile”
“Subscribers”, billions [14]
5.3
7.2
7.9
“Mobile”
Electricity use, TWh [17]
204
152
136
“Mobile”
Electricity use, TWh [18]
70
117
138
“Data Centers”
Data traffic, EB [7]
1403
4803
13761
“Data Centers”
Data traffic, EB [13]
1403
4803
17510
“Data Centers”
“Servers”, millions [15]
38
43
48
“Data Centers”
Electricity use, TWh [17]
196
220
207
“Data Centers”
Electricity use, TWh [18]
273
245
231
“Data Centers”
Electricity use, TWh [19] “Sobriety scenario”
227 (323 in 2013)
400
651
“Fixed”+ “Mobile” + “Data Centers”
Data traffic, EB [13]
1403
4803
17510
“Fixed”+ “Mobile” + “Data Centers”
Electricity use, TWh [17]
604
552
514
“Fixed”+ “Mobile” + “Data Centers”
Electricity use, TWh [18]
423
447
459
“Fixed”+ “Mobile” + “Data Centers”
Electricity use, TWh [19] “Sobriety scenario”
624 (757 in 2013)
863
1227
“Fixed”+ “Mobile” + “Data Centers”
Capita, billions [7]
6.85
7.04
7.4
“Fixed”+ “Mobile” + “Data Centers”
Total global instructions, nonillions (1030), (Tables 7
and 8)
0.15
1.04
7.7
Table 3: CAGR Values for Activity Metrics and Electricity Intensities in and between 2025 and 2030
Network type
E (TWh) and AM
Assumed
CAGRAM, 2020 to2030
2025 result
2030 result
“Fixed”
Data traffic, ExaByte (EB) [7]
26.6%
7693
25901
“Fixed”
Electricity use, TWh [17]
-13.2%. Nonlinear improvement of TWh/EB
from -20% in 2022
to -5% in 2030
204
448
“Fixed”
Data traffic, EB [13]
23.43% [13]
7357
21077
“Fixed”
Electricity use, TWh
-18.2% TWh/EB [EI2010 to 2020]
179
188
“Fixed”
“Subscribers”, billions [14]
1.5% [13]
2.1
2.3
“Fixed”
Electricity use, TWh
-1% TWh/billion subscribers
177
183
“Fixed”
Data traffic, EB [13]
23.43% [13]
7357
21077
“Fixed”
Electricity use, TWh [18]
-17.81% TWh/EB [EI2010 to 2020]
97
104
“Mobile”
Data traffic, EB [17]
79%
9722
178324
“Mobile”
Electricity use, TWh. 5G 0.06
TWh/EB
-35.6% [Nonlinear improvement of sub TWh/EB
(e.g. 4G TWh/EB) from -20% in 2022
to -5% in 2030] [17]
168
369

54 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
(Table 3). Continued.
Network type
E (TWh) and AM
Assumed
CAGRAM, 2020 to2030
2025 result
2030 result
“Mobile”
Data traffic, EB [13]
45%
3157
20257
“Mobile”
Electricity use, TWh
-29.4% TWh/EB (2010 to 2020 CAGR)
153
172
“Mobile”
Data traffic, EB [17]
79%
9722
178324
“Mobile”
Electricity use, TWh
-29.4% TWh/EB (2010 to 2020 CAGR)
439
1413
“Mobile”
“Subscribers”, billions
1.25%
8.64
9.5
“Mobile”
Electricity use, TWh
-8.72% TWh/billion subscribers
92
62
“Mobile”
Data traffic, EB [13]
45%
3157
20257
“Mobile”
Electricity use, TWh [18]
-21.36% TWh/EB (2010 to 2020 CAGR)
266
514
“Data Centers”
Data traffic, EB [7]
33.8%
43748
254498
“Data Centers”
Electricity use, TWh [17]
-12.4%. Nonlinear improvement of TWh/EB
from -20% in 2022 to -5% in 2030
249
944
“Data Centers”
Data traffic, EB [13]
25%
53438
163078
“Data Centers”
Electricity use, TWh
-22% TWh/EB (2010 to 2020 CAGR)
183
163
“Data Centers”
Data traffic, EB [22]
56%
161778
1494659
“Data Centers”
Electricity use, TWh
-22% TWh/EB (2010 to 2020 CAGR)
556
1495
“Data Centers”
“Servers”, millions [15]
1.4%
52
55
“Data Centers”
Electricity use, TWh
-1.77% TWh/ million servers
205
198
“Data Centers”
Data traffic, EB [13]
25%
53438
163078
“Data Centers”
Electricity use, TWh based on
[18]
-24% TWh/EB (2010 to 2020 CAGR)
184
146
“Data Centers”
Data traffic, EB [13]
25%
53438
163078
“Data Centers”
Electricity use, TWh based on
[19]
-14% TWh/EB (2010 to 2020 CAGR)
952
1393
Combined
“Fixed”+ “Mobile” +
“Data Centers”
Data traffic, EB [17]
33.8%
43748
254498
“Fixed”+ “Mobile” +
“Data Centers”
Electricity use, TWh
-15.5% TWh/EB (2010 to 2020 CAGR)
622
1761
“Fixed”+ “Mobile” +
“Data Centers”
Data traffic, EB [13]
25% [13]
53438
163078
“Fixed”+ “Mobile” +
“Data Centers”
Electricity use, TWh
-24% TWh/EB (2010 to 2020 CAGR)
410
326
“Fixed”+ “Mobile” +
“Data Centers”
Data traffic, EB [22]
56%
161778
1494659
“Fixed”+ “Mobile” +
“Data Centers”
Electricity use, TWh
-24% TWh/EB (2010 to 2020 CAGR)
1240
2991
“Fixed”+ “Mobile” +
“Data Centers”
Data traffic, EB [13]
25% [13]
53438
163078
“Fixed”+ “Mobile” +
“Data Centers”
Electricity use, TWh
-17% TWh/EB (2010 to 2020 CAGR)
1486
1800
“Fixed”+ “Mobile” +
“Data Centers”
Data traffic, EB [13]
25% [13]
53438
163078
“Fixed”+ “Mobile” +
“Data Centers”
Electricity use, TWh, based on
[18]
-22% TWh/EB (2010 to 2020 CAGR)
413
472

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 55
Table 3 shows that raising the number of “mobile”
and “fixed” subscribers by an order of magnitude -
which is likely for IoT and other kinds of
subscriptions/connections which will be very different
from current telecom subscribers - would increase the
“fixed” and or “mobile” TWh by an order of magnitude.
Data generated by IoT depends on the application.
Data traffic is more credible than subscribers for
predicting “fixed” and “mobile” networks electricity use,
while “servers” and Global IP Data Center traffic [13]
are well aligned whenever 2010 to 2020 electricity
improvement continue between 2020 and 2030 for data
traffic, i.e.
EI2010 to 2020 = EI2020 to 2030.
While having a tendency to underestimate e.g.
mobile traffic, [13] is a transparent basis for simplistic
high-level predictions of ICT networks electricity use.
Based on Tables 2 and 3, Figures 1 to 4 show a
graphical summary of the spread of electricity use
between 2010 and 2030 obtained by high-level
simplistic trend analyses.
4. DISCUSSION
2030 is approaching fast and earlier data traffic
based approaches [7,17] might have overestimated the
TWh of networks and data centers in 2025 and 2030 as
the historical 2010-2020 trend was not assumed to
continue.
Figure 1: Spread of electricity use for fixed networks 2010-2030.
Figure 2: Spread of electricity use for “mobile” networks 2010-2030.

56 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
Memristors (“memory+transistor”) might be one of
the game changers that will keep the electricity
improvements (EI) going 2020 to 2030 [23].
For data centers, if the current developments
continue, Hintemann and Hinterholzer argued that the
energy consumption of data centers will “only” double
by 2030 compared to today [24]. That estimation is
more consistent with a waning EI than constant EI
between 2020 and 2030. As shown in Figure 3, the
current data center electricity is quite uncertain [19,20]
and a higher starting value in 2020 results in higher
electricity use in 2030.
Also, there are estimations of 79% CAGR of mobile
data traffic [17] from 2020 to 2030. Using 79% - instead
of 45% [13] - with constant EI will give >1400 TWh for
mobile networks in 2030, instead of 172 TWh. This
shows that a more complex model [7] for mobile can be
less sensitive to the total traffic and that the total
electricity use is very sensitive to data traffic growth in
simplistic modelling.
In reality there are many layers of intensity and
efficiency at sector, sub-network, and site equipment
level.
The convergence of fixed and mobile networks
might speak against using separate traffic data, but this
is doubtful as long as sources like [13] continue to
predict separate network traffic data.
Figure 3: Spread of electricity use for “data centers” 2010-2030.
Figure 4: Spread of electricity use for combined “mobile”, “fixed”, “data center” networks 201+-2030.

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 57
Increased electricity use in the use stage of
Networks and Data Centers 2020 to 2030 may be
driven by decelerating energy efficiency.
The present research shows that the data traffic
modelling approaches are not wrong per se, and that
the assumptions of gradually slowing EI improvement
from 2022 – used within the data traffic approaches -
are realistic. However, waning Moore´s law - and
transistor switching energy challenges [12] - will be
mitigated by “smart” engineering and system level
design of networks and data centers? On the other
hand the demand for ultra-low times required for
transmitting a message through the network – i.e. ultra-
low latency [25] - could mitigate the trend of building
fewer large hyperscale data centers, in turn believed to
offer huge overall power saving opportunities. The
fundamental conflict in ICT is between capacity, latency
and energy efficiency. Moreover for mobile networks it
might not be enough to phase out older equipment and
switch to “5G” equipment as all spectrums have to be
offered simultaneously and each spectrum (e.g.
3.5GHz band) has its own energy efficiency [26].
Additionally the first generation of “5G” equipment
might not be optimized as far as power use [31].
Server shipments could well grow faster between
2020 and 2030 than the preceding decade. Servers are
also very different. Fuchs et al. [27] found that idle
server power demand to be significantly higher than
benchmarks from ENERGY STAR and the industry-
released SPEC database, and SPEC server
configurations—and likely their power scaling—to be
atypical of volume servers.
Further Diouani and Medromi proposed simplified
formulae for energy consumption estimation of cloud
data centers [28]. However, no quantification of global
energy use of cloud data centers was demonstrated.
Continuation of -22% TWh/EB annual improvement
for data centers as a whole (Table 3) - and for the
servers within - between 2020 and 2030 will require a
large percentage of hyperscale data centers normally
offering higher utilization rate and power use
effectiveness including liquid cooling and similar [29].
4.1. “Fixed” + “Mobile”+ “Data Centers” – can the
Fixed Mobile Convergence be Forecasted?
Table 4 shows Extrapolation with other approaches
from 2020 baseline values to 2025 and 2030 for
Networks and Data Centers.
As shown in Table 4, alternative extrapolation
approaches give quite different results for 2030. There
will be a manageable increase of the electric power,
but according to the scientific method the high-level
method can only assume that historically achieved
electricity intensities - and data traffic changes - will
continue until new observations have been done. It has
never been argued by data traffic proponents that the
electricity use (TWh) rises at the same rate as the
ExaBytes.
The developing world will develop their
infrastructure and a thought experiment can include
network and data center TWh/capita extrapolation to
the World capita.
The high-level approaches – especially those
combining all networks and data centers - discussed in
the present research likely underestimate the total
TWh. Using total global instructions for AM give
somewhat unrealistic forecasts when based on
Table 4: Extrapolation with other Approaches from 2020 Baseline Values to 2025 and 2030 – Networks and Data
Centers
Network type
TWh and Data traffic
CAGR
2025
2030
“Data Centers”
Historical Compound Aggregated Growth
Rate (CAGR) of electricity TWh !TWh
4.4% [16]
256
318
“Fixed”+”Mobile”
Historical Compound Aggregated Growth
Rate (CAGR) TWh !TWh
10.4% [16]
503
825
“Fixed”+ “Mobile” + “Data Centers”
Historical Compound Aggregated Growth
Rate (CAGR) TWh !TWh
759
1143
“Fixed”+ “Mobile” + “Data Centers”
Global Capita, billions
1.5%
8
8.6
“Fixed”+ “Mobile” + “Data Centers”
TWh/Global Capita
-2%
493
471
Fixed”+ “Mobile” + “Data Centers”
Total global instructions, nonillions
44% (Tables 7 and 8)
48
298
“Fixed”+ “Mobile” + “Data Centers”
TWh/nonillion instructions
-33.75% for TWh/nonillion bits
402
328

58 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
historical trends as demonstrated by 402 and 318 TWh
in 2025 and 2030, respectively. Similar results are
obtained by using [13] total global data center IP traffic
(Table 3, Combined). The main reason is that each
network and data center type has its own energy
efficiency characteristics which are lost in calculation
which combine all types into one score.
4.2. “Fixed”+”Mobile”
Van Heddeghem et al. [16] proposed growth rate for
electricity use observed by 2007 to 2012 may not have
continued until 2020. Therefore it is questionable if
those growth rates - 10.4% for networks and 4.4% for
data centers - can be used for 2020 to 2030.
4.3. Metrics for Data Center Activities – can one
Activitry Metric be Used as Proxy for Others?
Do data center activities based on i) data storage, ii)
data traffic and iii) computations, track each other
consistently so that one of these could be used as a
proxy for all three?
Tables 5 and 6 show that Global data center IP
traffic [13] and communication volume [30] are well
suited for acting as proxy for other metrics.
Data center IP traffic is expected grow at a
Compound Annual Growth Rate (CAGR) of 25 percent
from 2016 to 2021 [13]. Overall data center workloads
and compute instances are predicted to more than
double (2.3-fold) from 2016 to 2021 [13]. The growth
rate of workloads seems to be slower than the data
traffic. Xu’s [30] instruction/s growth rate and Cisco’s
compute instances growth rate [13] are not consistent.
The historical improvements may continue and it
looks much better for data centers than previously
estimated thanks to heavy focus on energy efficiency
research and implementation. The awareness is strong
and the engineers are smart. However, the 2020 global
baseline for data centers is much more than 200 TWh if
[19,20] are accurate.
4.4. Rationale for Data Traffic Method Using
Published Estimations
There are several reasons why data traffic is a
better activity metric for forecasting electricity use of
ICT Networks. Here follows four main reasons:
1. Personalized usage profile for same subscriber
type.
How can a user making phone a phone call be
compared to a user watching 4K video? Both these
Table 5: Data Center Activities Expressed as Data Storage, Comm unication Volume and Instructions Per Second [30]
Year
Global Storage volume [30]
Global Communication volume [30]
Global Instructions per second [30]
1986
21 PetaByte(PB)
59 PB
0.74 Peta
2007
277 ExaByte (EB)
537 EB
195 Exa
2010
2015
2020
2025
31.2% CAGR between 2007 and
2030
31.2% CAGR between 2007 and 2030
51% CAGR between 2007 and 2030
2030
140 ZettaByte (ZB)
272 ZB
2588 Zetta
Table 6: Data Center Activities Expressed as Data Storage, Data Traffic and Workloads [13]
Year
Storage volume [13]
Communication
volume
Global Data center IP traffic
[13]
Overall data center workloads and
compute i nstan ces [13]
1986
?
2007
?
2010
1.37 ZB
2015
600 EB
5.4 ZB
1
2020
2.6 ZB
16.6 ZB
2.3
2025
32% CAGR between
2016 and 2021
25% CAGR between 2016 and
2021
18% CAGR between 2016 and 2021
2030

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 59
users can be the same type of subscriber (follow the
same subscriber package with an operator), but can
they both declare the same network electricity
consumption?
2. Human and other types of subscribers
A company could be a subscriber of public cloud. A
vehicle could be a subscriber. How do they compare
with a human subscriber? There will be many different
types of subscribers – and more quantity - in the future
than humans. A company having one fixed broadband
service (subscription) has several fixed users of the
same service.
3. Communication era is changing to big data era
Data processing will increase faster than data
transportation. For legacy networks, the access
networks is the biggest part of the electricity
consumption and that in turn is well correlated to
“subscribers”. However, for the big data era, data
processing and storage will be higher than traditional
telecom, i.e. ICT and not only communication.
4. Lack of dependence between power and
subscribers in the 5G era
AMs like the area covered, the number of
subscribers, or the amount of data traffic have had a
small impact on the amount of electricity used by 1G,
2G, 3G, and 4G mobile networks. Frenger and Tano
argued that 5G could change this and make the
network energy much more proportional to the actual
network load (=amount of data) [31].
The number of subscribers will reach a limit, and
therefore data per capita is probably the best AM if
data and human subscribers are to be combined.
4.5. Computing Share of Total Primary Energy
Supply
The rationale for the calculating the total global
number of bits (instructions) is not clear. Cisco’s total
ZettaByte data traffic served will - in itself if translated
to bits - render too few bits to result in 8% as share of
computing if multiplied with e.g. 5 pJ/bit. Barlage
mentioned 7-10% of computing of current total PES
and current 3 pJ/computation [9]. The amount of data
will increase different amount of times between 2020
and 2030 depending on which measure is used (data
volume, global data center IP traffic, instructions,
logical operations, computations).
In summary:
1. Data traffic and data traffic intensity will be much
more related to energy usage of networks and
data centers than subscribers.
2. Data traffic can more sophistically forecast future
energy usage levels than subscribers.
3. Data traffic work well at sub-sector level and
data [13] are available.
4. The development of data traffic intensity
roadmaps will work well on company level.
5. Data traffic and data traffic intensity evolution is
applicable and relevant to all companies in the
sub-sector.
6. Data traffic and data traffic intensity is adequate
for past, present and future development.
7. Data traffic and data traffic intensity respond very
well – much better than subscribers - to
significant changes in technology and its
deployment.
8. Data traffic is very easy to understand and
interpret.
4.6. Roadmaps for J/Instruction
Tables 7 and 8 show a new model for estimation of
the primary energy use of computing.
Table 7 shows that instructions/byte increases 15%
per year. Table 8 shows that the total global
instructions might increase 44% per year and the
shares of computing of primary energy supply (PES)
considering various roadmaps for J/instruction [8].
The instructions in Table 8 is obtained by
multiplying instructions/byte [30] in Table 7 with Cisco
Global IP Data Center Traffic [13], 17.1 ZettaByte
(expressed as 1.6×1023! bits)!in 2020 and 159 ZettaByte
(expressed as 1.5×1024 bits). Miller [33] argued that we
should move from energies for ∼1 cm to∼10 m
interconnects that are currently in the range of pJ (or
larger) total energy per bit, down towards ∼10 fJ or
lower total energy per bit. This argumentation is
reflected in Table 8. Nevertheless, the ZettaBytes
might be much higher in 2030 [22].
Data centers have and will probably develop along
the best case scenario in [7]. Best case, Figure 4 in [7]

60 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
is the most probable for data centers as also suggested
in Figure 3 by three different calculations.
Estimation results from somewhat more
comprehensive models than those investigated in the
present research – made with some updates of [7,17] -
for mobile networks and others are shown in Table 9.
As shown in Table 3, the TWh are much lower if no
waning of Moore’s law is factored in. Also a new
assumption is that 5G – starting at 0.06 TWh/ExaByte
[7] - will not use historical improvement of the TWh/EB
factor between 2010 and 2020 as assumed by [7].
2030 5G TWh are calculated as: 12 months×94%
share 5G×2537 EB/month×0.06 TWh/EB×0.229
(accumulated improvement factor in 2030 with waning
Moore’s law. Compare the 0.229 accumulated
improvement factor to 0.7810=0.083 with no waning and
0.053 in [7]) = 393 TWh.
4.7. Solutions to Counteract the Increases in Global
Energy and Electricity Usage in ICT Networks and
Data Centers
There are some ways in which the electricity use of
networks and data center can be reduced. Suggestions
are energy efficient software coding, neural networks
which mimic the human brain and moving the memory
storage closer to the computation [9]. Others are timers
on Wi-Fi modems shutting them down during night time
Table 7: Increase of Instructions Generated Per Da ta Volume 2007 to 2030
Xu Instructions/year [30]
Xu Data volume, Byte/year [30]
Xu Instructions/Xu Byte
2007
7.1×1027
9.38×1020
7.55×106
2010
2.4×1028
2.12×1021
1.16×107
2015
1.9×1029
8.25×1021
2.35×106
2020
3.5×1030
5.53×1022
4.79×107
2025
1.2×1031
1.24×1023
9.74×107
2030
9.6×1031
4.86×1023
1.98×108
Table 8: Estimation of Primary Energy Supply of Computing from Total Global Instructions and Energy Per
Instruction
Total global
Instructions/year = Xu
Instructions/ Xu Byte ×
Global IP data center
traffic [13]
J/instruction
[8]
J
EJ
Total Primary
Energy Supply
(PES), EJ [1, 32]
Share Computing
of PES
2020, 5
pJ/instruction (i)
4.79×107 (Table
7)×17510(Table 2)×260 ×8
bits/Byte=7.7×1030
5×10-12
3.9×1019
38.5
600
6.4%
2030, 5 pJ/i
3.0×1032
5×10-12
1.5×1021
1500
790
189%
2030, 0.3 pJ/i
3.0×1032
3×10-13
9.6×1019
90
790
11%
2030, 10fJ/i
3.0×1032
1×10-14
3.2×1018
3
790
0.38%
2030, 10aJ/i
3.0×1032
1×10-17
3.2×1015
0.003
790
≈0%
2030, 3zJ/i
3.0×1032
3×10-21
9.6×1011
≈0
≈0%
Table 9: Forecasts of Electricity Use (TWh) of the Entire Internet Including ICT Networks 2020 to 2030
“Fixed”
networks
use
“Mobile”
networks
use
“Data
centers”
use
Consumer devices
use
incl. TV&TV
peripherals
“Wi-Fi” use
Manufacturing of ICT
TOTAL
2020
171
98
299
966
72
382
1988
2025
200
116
412
918
99
304
2049
2030
428
446
974
839
234
313
3234

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 61
[34]. Others are to phase out older energy inefficient
fixed networks equipment [18] and optimize the data
use itself [35].
4.8. Sensitivity Analyses
The sensitivities
dE
dEI2010 to 2020
!
"
#
#
$
%
&
&
and
dE
CAGRAM, 2020 to 2030
!
"
#
#
$
%
&
&
are tested by increasing
EI2010 to 2020
and
CAGRAM, 2020 to 2030
1% and determine how
!
changes in 2025 and 2030. Mobile networks in Table 2
and 3 – with
CAGRAM, 2020 to 2030 =45%
and
EI2010 to 2020 =29.4%
are chosen. The baseline values
are
E2025 =153 TWh
and
E2030 =172 TWh
. When
CAGRAM, 2020 to 2030
is changed to 46.45% the following is
obtained:
E2025 =161 TWh
and
E2030 =190 TWh,
i.e.
dE
CAGRAM, 2020 to 2030
!
"
#
#
$
%
&
&
is +5.2% for 2025 and +10.4% for
2030.
When
EI2020 to 2030
is changed to 29.7% the following
is obtained
E2025 =150 TWh
and
E2030 =160 TWh,
i.e.
dE
CAGRAM, 2020 to 2030
!
"
#
#
$
%
&
&
is -2% for 2025 and -7% for 2030.
This suggests that the resulting TWh is more
sensitive to traffic growth than electricity improvement.
5. CONCLUSIONS
The conclusions for top-down prediction methods
for ICT networks and data centers global electricity use
are:
• Data traffic numbers from [13] are – if used
carefully - the most reliable prediction bases for
networks and data centers electricity use.
• TWh for all kind of networks is not equally
sensitive to changes in Data traffic growth and
electricity intensity improvement.
In summary both hypotheses set up in Section 1
could not be falsified. The present research has proven
that using data (traffic) – being close to an SI unit – can
give robust and reasonable prediction results for ICT
Networks and data center electricity use. There is not a
precision problem with historical [7] data based
prediction approaches, however the actual change
potential of the electricity intensity has not been
implemented carefully enough. It cannot be ruled out
that the electricity intensity improvement will continue in
the next decade thanks to smart engineering, and then
there is no issue of power costs. Subscriber based AM
worked well historically but is not well-suited for an
uncertain future in which the main certain fact is that
data traffic will increase heavily. Capita based AM can
only give order of magnitude indications, but are too
crude for well-founded predictions. Global data center
IP traffic [13] can be used as a proxy for other AM.
Data traffic numbers [13] (having relatively low
uncertainty) and historical TWh/Exabyte improvement
numbers are the best data known in this field. How
these numbers might deviate between 2020 and 2030
is to be the key discussion.
The main argument of this research is that a data
traffic approach does indeed give very reasonable
numbers for the main global ICT Networks and data
centers energy and electricity evolution.
6. NEXT STEPS
Each model has very few factors compared to a
rather complex reality. However, measuring ICT power
consumption is a hands-on problem which might
involve smart meters.
The relative results (Tables 3 and 4) might look very
different to each other suggesting that simplistic
approaches have challenges. The next step is to
translate the top-down electricity intensity
improvements to product equipment targets. The top-
down EI improvements are more suitable for network
operator targets setting than equipment manufacturer
roadmaps. The first assumption is that equipment –
e.g. servers and base stations – are expected to have
the same roadmap as the data centers and mobile
networks of which they are parts. “Billion shipped
memories and processors” and “billion connections”
are worthwhile for AM analysis.
Scaling up the effect of AI training and the shifting
tide towards edge computing on total electricity use is
in the cards.
ACKNOWLEDGMENTS
Anonymous reviewers are greatly appreciated for
comments, which improved this paper. Zhu Bin, Tomas
Edler, Avelino Benavides, Ulrik Imberg and Magnus
Olsson are acknowledged for valuable comments.

62 International Journal of Green Technology, 2019, Vol. 5 Anders S.G. Andrae
AUTHOR CONTRIBUTIONS
Anders S. G. Andrae wrote the paper.
CONFLICTS OF INTEREST
The author declares no conflict of interest. The
views of this paper is the authors own and not those of
the company.
REFERENCES
[1] BP Statistical Review of World Energy. June 2019. 68th
Edition. [cited 2019 Sept 25]: Available from:
https://www.bp.com/en/global/corporate/energy-
economics/statistical-review-of-world-energy.html
[2] Dyson F. The need for sustainable heretics. In: Madhavan G,
Oakley B, Green D, Koon D, Low P, (Eds.) Practicing
Sustainability. Springer, New York, NY, 2013; 71-76. 2013.
https://doi.org/10.1007/978-1-4614-4349-0_14
[3] Lee JY, Kim H. Ambient air pollution-induced health risk for
children worldwide. The Lancet Planetary health 2018; 2,
e285-e286.
https://doi.org/10.1016/S2542-5196(18)30149-9
[4] Wikipedia. Energy efficiency in transport. [cited 2019 Sept
25]: Available from: https://en.wikipedia.or g/wiki/
Energy_efficiency_in_transport
[5] Romare M, Dahllöf L. The life cycle energy consumption and
greenhouse gas emissions from lithium-ion batteries.
Stockholm. [cited 2019 Sept 25]: Available from:
http://www.ivl.se/download/18.5922281715bdaebede95a9/14
96136143435/C243.pdf
[6] Kalghatgi G. Is it really the end of internal combustion
engines and petroleum in transport?. Appl Ener 2018, 225,
965-74.
https://doi.org/10.1016/j.apenergy.2018.05.076
[7] Andrae ASG, Edler T. On global electricity usage of
communication technology: trends to 2030. Challenges 2015;
6: 117-57.
https://doi.org/10.3390/challe6010117
[8] Åberg M, Mämmelä A. End of Moore’s law: performance
limited to energy limited microelectronics. 2016. [cited 2019
Sept 25]: Available from: https://www.vttresearch.com/
Documents/Eemeli/21st%20Eemeli/15%20Åberg%20Eemeli
21_MÅ_03_16.pdf
[9] Barlage D. Impact of the Digital Revolution on World Wide
Energy Consumption. [cited 2019 Sept 25]: Available from:
https://aroundtheworld.ualberta.ca/portfolio/douglas-barlage/
[10] Techxplore. Light: A possible solution for a sustainable AI.
[cited 2019 Sept 25]: Available from: https://techxplore.com/
news/2019-07-solution-sustainable-ai.html
[11] MIT Technology R eview. Training a single AI model can emit
as much carbon as five cars in their lifetimes. [cited 2019
Sept 25]: Available from: https://www.technologyreview.com/
s/613630/training-a-single-ai-model-can-emit-as-much-
carbon-as-five-cars-in-their-lifetimes
[12] Andrae ASG. Prediction Studies of Electricity Use of Global
Computing in 2030. Int J Sci Eng Invest 2019; 8: 27-33.
[cited 2019 Sept 25]: Available from: http://www.ijsei.com/
papers/ijsei-88619-04.pdf
[13] Cisco Global Cloud Index. November 19, 2018. [cited 2019
Sept 25]: Available from: https://www.cisco.com/c/en/us/
solutions/collateral/service-provider/global-cloud-index-
gci/white-paper-c11-738085.html#_Toc503317524
[14] International Telecommunication Union. [cited 2019 Sept 25]:
Available from: https://www.itu.int/en/ITU-D/Statistics/Pages/
stat/default.aspx
[15] Statista. Global Server Shipments by Vendor. [cited 2019
Sept 25]: Available from: https://www.statista.com/statistics/
267390/global-server-shipments-by-vendor
[16] Van Heddeghem W, Lambert S, Lannoo B, Colle D, Pickavet
M, Demeester P. Trends in worldwide ICT electricity
consumption from 2007 to 2012. Comput Commun 2014; 50:
64-76.
https://doi.org/10.1016/j.comcom.2014.02.008
[17] Andrae ASG. Projecting the chiaroscuro of the electricity use
of communication and computing from 2018 to 2030. [cited
2019 Sept 25]: Available from: https://www.researchgate.
net/publication/331047520_Projecting_the_chiaroscuro_of_t
he_electricity_use_of_communication_and_computing_fr om
_2018_to_2030
[18] Malmodin J, Lundén D. The energy and carbon footprint of
the global ICT and E&M sectors 2010–2015. Sustainability
2018; 10(9): 3027.
https://doi.org/10.3390/su10093027
[19] The Shift Project. Lean ICT – Towards Digital Sobriety. [cited
2019 Sept 26]: Available from: https://theshiftproject.org/wp-
content/uploads/2019/03/Lean-ICT-Report_The-Shift-
Project_2019.pdf
[20] Computer World. [cited 2019 Sept 25]: Available from:
https://www.computerworld.com.au/article/666183/power-
consumption-china-internet-usage-grow-two-thirds-by-2023/
[21] Kyriakopoulos CA, Nicopolitidis P, Papadimitriou GI,
Varvarigos E. Fast Energy-Efficient Design in Elastic Optical
Networks Based on Signal Overlap. IEEE Access 2019; 7:
113931-41.
https://doi.org/10.1109/ACCESS.2019.2935328
[22] Shi W, Pallis G, Xu ZW. Edge Computing [Scanning the
Issue]. Proc IEEE 2019; 107(8): 1474-81.
https://doi.org/10.1109/JPROC.2019.2928287
[23] Gnawali KP, Mozaffari SN, Tragoudas S. Low Power Artificial
Neural Network Architecture. [cited 2019 Sept 25]: Available
from: https://arxiv.org/abs/1904.02183
[24] Hintemann R, Hinterholzer S. Energy consumption of data
centers worldwide. [cited 2019 Sept 25]: Available from:
http://ceur-ws.org/Vol-2382/ICT4S2019_paper_16.pdf
[25] Chen H. Abbas R, Cheng P et al. Ultra-reliable low latency
cellular networks: Use cases, challenges and approaches.
IEEE Commun Mag 2018; 56(12): 119-125.
https://doi.org/10.1109/MCOM.2018.1701178
[26] Huo L, Jiang D, Lv Z. Soft frequency reuse-based
optimization algorithm for energy efficiency of multi-cell
networks. Comput Elect Eng 2018; 66, 316-31.
https://doi.org/10.1016/j.compeleceng.2017.09.009
[27] Fuchs H, Shehabi A, Ganeshalingam M, Desroches LB, Lim
B, Roth K, Tsao A. Comparing datasets of volume servers to
illuminate their energy use in data centers. Energ Effic 2019;
1-14.
https://doi.org/10.1007/s12053-019-09809-8
[28] Diouani S, Medromi, H. How energy consumption in the
cloud data center is calculated. Proceedings of International
Conference of Computer Science and Renewable Energies
(ICCSRE); 2019: Agadir, Morocco: IEEE; 2019.
https://doi.org/10.1109/ICCSRE.2019.8807458
[29] Nadjahi C, Louahlia H, Lemasson S. A review of thermal
management and innovative cooling strategies for data
center. Sustain Comput Inform Systems 2018; 19: 14-28.
https://doi.org/10.1016/j.suscom.2018.05.002
[30] Xu ZW. Cloud-sea computing systems: Towards thousand-
fold improvement in performance per watt for the coming
zettabyte era. J Comput Sci Technol 2014; 29(2): 177181.
https://doi.org/10.1007/s11390-014-1420-2
[31] Frenger P, Tano R. More Capacity and Less Power: How 5G
NR Can Reduce Network Energy Consumption. Proceedings
of 2019 IEEE 89th Vehicular Technology Conference

Comparison of Several Simplistic High-Level Approaches International Journal of Green Techno logy, 2019, Vol. 5 63
(VTC2019-Spring); 2019: Kuala Lumpur, Malaysia: IEEE:
2019: p. 1-5.
https://doi.org/10.1109/VTCSpring.2019.8746600
[32] US Energy Information Administration. EIA projects 28%
increase in world energy use by 2040. [cited 2019 Sept 25]:
Available from: https://www.eia.gov/todayinenergy/
detail.php?id=32912
[33] Miller DA. Attojoule optoelectronics for low-energy
information processing and communications. J Lightwave
Technol 2017; 35(3): 346-96.
https://doi.org/10.1109/JLT.2017.2647779
[34] Terry N, Palmer J. Trends in home computing and energy
demand. Build Res Inform 2015; 44(2): 175-87.
https://doi.org/10.1080/09613218.2015.1040284
[35] Morley J, Widdicks K, Hazas M. Digitalisation, energy and
data demand: The impact of Internet traffic on overall and
peak electricity consumption. Energ Res Soc Sci 2018; 38:
128-37.
https://doi.org/10.1016/j.erss.2018.01.018
Received on 25-08-2019 Accepted on 28-09-2019 Published on 02-10-2019
DOI: https://doi.org/10.30634/2414-2077.2019.05.06
© 2019 Anders S.G. Andrae; International Journal of Green Technology
This is an open access article licensed under the terms of the Creative Commons Attribution Non-Commercial License
(http://creativecommons.org/licenses/by-nc/3.0/) which permits unrestricted, non-commercial use, distribution and reproduction in
any medium, provided the work is properly cited.